The Retraining Trap

Why $500M Won’t Fix the AI Transition (and What Actually Will)

Why $500M Won’t Fix the AI Transition (and What Actually Will)

In June 2026, a coalition of the world’s largest AI companies announced Raise Us, a $500M nonprofit initiative to fund state-level retraining programs for workers displaced by artificial intelligence. The names behind it read like a who’s who of the AI industry: OpenAI, Anthropic, Microsoft, Amazon, Bank of America, Workday, IBM, among others. Former Commerce Secretary Gina Raimondo is co-leading it, alongside former Indiana governor Eric Holcomb. This is a good sign. The companies building AI are acknowledging that the transition creates real costs for real people. That acknowledgment has been a long time coming. The scale of the commitment signals they understand those costs are serious. But acknowledgment and correct diagnosis are different things. The instinct driving Raise Us is that the right response to AI workforce displacement is to retrain workers in new skills. That instinct is almost certainly wrong. Not because the workers don’t need support. Because skills aren’t the bottleneck.

The Ceiling We Keep Building

State-level workforce retraining programs have a documented history of underperformance. The evidence on their effectiveness is not encouraging. Gina Raimondo knows this. During her tenure as governor of Rhode Island, she ran one of the more ambitious state retraining initiatives in the country. In launching Raise Us, she has said past federal retraining efforts were largely “ineffective.” That assessment wasn’t unfair. The research on workforce retraining consistently finds the same pattern. Evaluations of federal programs including JTPA and Trade Adjustment Assistance found no statistically significant employment or earnings gains for participants, and TAA participants remained underemployed relative to non-TAA workers even four years after displacement (Brookings: Jacobs, “AI Labor Displacement and the Limits of Worker Retraining,” May 2025). A 2026 analysis of 23 million WIOA training records found the program rarely shifted participants into less automation-exposed work (arXiv 2605.03767). None of this is a knock on the people who design or deliver these programs. The problem is structural. Retraining programs operate on a model: there is a skill deficit, here is a curriculum, here is a credential. Complete the program, earn the credential, get the job. It’s a clean model. It just doesn’t reflect how people actually develop competency in new work. Skills are learned in context. They stick when practiced on real problems, in real teams, under real conditions. The classroom gets you started. The room is where you actually learn. The other problem is more fundamental than program design. For the AI transition specifically, the premise is wrong. Individual skills are not the primary obstacle to a successful transition. They are not where the work stalls.

The Bottleneck Nobody Is Naming

Earlier this year, I spoke with the head of AI enablement at a major North American bank. He had rolled out AI tools to over 4,000 employees. Training completions were high. His team had built a solid onboarding curriculum. Adoption numbers looked reasonable on paper. But the work wasn’t changing. Teams were using AI to produce the same deliverables faster. Nobody was questioning whether the deliverables still needed to exist. Nobody was renegotiating who reviewed what. Senior analysts were reviewing AI-generated work at the same rate they had reviewed junior analyst work, which defeated the efficiency argument entirely. The tools were in the hands of the individuals. The operating model of the teams hadn’t moved. He didn’t have a skills problem. His employees could use the tools. What they couldn’t do was change how they worked together around those tools. This is the transition that nobody is naming. It isn’t a skills transition. It is a delegation transition. When AI can handle tasks that used to require human time and judgment, every team has to renegotiate who decides what, who checks what, and who is accountable for what. An individual contributor who used to produce the first draft now manages the agent that produces the first draft. The work changes. The role changes. The team’s accountability structure has to change with it, or you end up with a faster version of the same old process. That renegotiation happens at the team level. It can’t be solved at the individual level. You can train a person to manage an AI agent. You can’t train a person to lead a team that has learned to work with AI agents together, because the team has to do that learning together. The bank’s head of AI enablement eventually figured this out. The intervention that moved the needle wasn’t another training program. It was a six-week sprint where teams worked on real deliverables, debriefed weekly on what they had tried and what had failed, and rebuilt their working agreements around AI as an actual participant in the workflow. Individual capability wasn’t the issue. Team operating model was. The sprint addressed the right unit.

Spark, Don’t Train

The distinction worth holding here is between training and sparking. Training is one-time, classroom-based, information-transfer. It works well for things that can be reduced to a curriculum: how to use a tool, how to follow a process, how to answer the questions on a certification exam. It does not work well for developing judgment, because judgment can’t be transferred from a teacher to a student. It has to be built through practice, feedback, and reflection on real conditions. Retraining programs are, by design, training programs. They are well-suited to the problem they are designed for. The AI transition is not that problem. The pattern that actually builds AI capability looks different. You model the behavior visibly. You practice it on real work. You reflect on what happened. Then you do it again. Not once. Repeatedly, over time, embedded in the actual work of the team. This is not a novel insight about how humans learn. It is how judgment develops in every high-stakes domain. Lawyers build judgment through cases. Surgeons build it in the operating room. Managers build it by managing under real conditions with real feedback. AI fluency follows the same pattern. The only way to build judgment about when to trust an AI output, when to override it, and how to calibrate your confidence in it is to use AI on real stakes, with real feedback, in real teams. There is a specific dimension of this that retraining programs are structurally unable to address. The judgment questions in AI adoption are not primarily individual. They are social. Who owns the output when an AI writes the first draft? How does a team validate an AI recommendation before acting on it? When a junior employee produces AI-assisted analysis faster than a senior employee can review it, whose judgment carries? These are not questions a certification exam can answer. They are questions a team has to work out together in real time. The organizations getting this right are designing for exactly this kind of learning. They are running team-level sprints on active projects where AI is embedded into the actual work. They are creating structured reflection time so that teams can process what’s working and what isn’t. They are building internal communities where practitioners share not just what the tools can do but when and how to use them well. They are putting leaders in the room as practitioners, not sponsors, because a leader who uses AI visibly gives a team permission to experiment with it openly. None of this shows up in a training completion report. All of it shows up in whether the organization can actually work differently twelve months from now.

The Move Is Organizational, Not Individual

This reframe has a direct implication for how the Raise Us funding should be deployed, and for how every organization should be thinking about their own AI transition work. The unit of change is not the individual worker. It is the team. Retraining programs that target workers one at a time, equip them with new credentials, and then return them to unchanged teams will produce unchanged outcomes. The ceiling is not the individual’s capability. It is the team’s operating model. This does not mean individual development is irrelevant. It means individual development cannot do the work alone. A newly credentialed worker returning to a team that has not redesigned how it makes decisions with AI is like a newly trained pilot returning to an airline that hasn’t updated its cockpit protocols. The individual is ready. The system isn’t. The system is what fails. The organizations navigating the transition well have understood something the retraining model misses: you redesign how teams work, and capability follows. You do not upgrade individuals and hope the team reconfigures around them. Redesigning how teams work means starting with the actual decisions a team makes and the actual workflows it runs, mapping where AI can and should participate, renegotiating the human roles around that participation, and then practicing the new model with real work. It is slower than training. It is harder to fund through traditional channels. It is what actually works. The Raise Us coalition has the resources to fund team-level capability building at scale. They could fund organizations to run AI adoption at the team level rather than the individual level. They could invest in practitioner communities where teams share what they’re learning in real time. They could fund the kind of embedded, real-work learning that the classroom model can’t reach. The money is not the problem. The model is.

The Transition Deserves Better

The workers displaced by AI deserve a serious response. The Raise Us commitment, whatever its design flaws, is a signal that the AI industry is taking the workforce question seriously. That signal matters. The conversation about AI and work has too often been led by people whose economic exposure to the transition is low. The fact that major AI companies are putting significant resources toward this is not nothing. But the workers affected by this transition don’t just need new skills. They need to work in organizations that have figured out how to work differently with AI. Individual retraining that deposits newly credentialed workers into unreconstructed teams is not a transition strategy. It is a deferral. Raimondo herself called the old model ineffective. The question is whether the new money is going toward a different model or the same one at larger scale. The right question to ask of Raise Us, and of every AI adoption initiative, is not how many workers are we retraining. It is how many teams are we rebuilding. That is a harder question to fund. It is also the right one. If you are designing an AI adoption program and wondering what team-level capability building actually looks like in practice, the Collaborative AI Lab is one place to explore it. The design principles are the same ones described here: model the behavior, practice on real work, reflect and rebuild. The transition deserves that kind of rigor. So do the people going through it.